Abstract: This paper proposes a gradient-similarity based multi-topic jointly pre-training method for automated essay scoring (AES). Specifically, in the pre-training stage, the training data of multiple topics are used at the same time, and the similarity between the gradient vector of a sample from other topics and the gradient vector of target topic is calculated as the loss weight for this sample. Besides, this paper also designs three types of handcrafted features, combining deep learning with feature engineering. Comparative experiments are conducted on publicly available datasets, and the results show that compared with the existing baselines, both proposed multi-topic jointly pre-training method and handcrafted features can effectively improve the scoring accuracy of the AES model.